Machine learning in neuroscience Bojan Mihaljevic, Luis Rodriguez-Lujan Computational Intelligence Group School of Computer Science, Technical University of Madrid 2015 IEEE Iberian Student Branch Congress April 24 th, Madrid B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 1 / 48
Outline 1 Introduction 2 Methods Machine learning Bayesian networks Directional statistics 3 Applications Introduction to neuroscience Neuron classification Morphological simulation Soma and spines DRCMST Other applications 4 Future work B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 2 / 48
Computational Intelligence Group At Artificial Intelligence Department, School of Computer Science Since 2008 2 full professors, 1 associate professor, 1 post-doc, and 11 PhD students B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 3 / 48
Outline 1 Introduction 2 Methods Machine learning Bayesian networks Directional statistics 3 Applications Introduction to neuroscience Neuron classification Morphological simulation Soma and spines DRCMST Other applications 4 Future work B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 4 / 48
A useful tool http://en.wikipedia.org/wiki/file:no-spam.png http://commons.wikimedia.org/wiki/file:logo_youtube.svg B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 5 / 48
Data-driven Learn from data what you cannot program (well) explicitly Large amounts of data these days Typically, we assume data comes as attribute values X 1 X 2 X 3 X 4 X 5 1.40 A 10003-24 D -0.31 B 2039 21 C -0.01 B 7383 70 U Goal: learn some function over X Related terms: data mining, pattern recognition, data science... B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 6 / 48
Tasks Classification (discrete target variable) Regression (real-valued target variable) Clustering (hidden discrete target variable) Others: collaborative filtering, market basket analysis, etc. http://commons.wikimedia.org/ wiki/file:social_red.jpg B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 7 / 48
Multiple models Some of them: P(x, c) = P(c)P(x 1 c)p(x 2 c)p(x 3 c)p(x 4 c)p(x 5 c) Naive Bayes k nearest neighbors p(c x, w) = Ber(y sigm(x T x)) Logistic regression Hastie, T., Tibshirani, R., Friedman, J., (2009). The elements of statistical learning (Vol. 2, No. 1). New York: Springer Decision tree B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 8 / 48
Toolbox Many different tools to extract models from data Optimization (often heuristic) Combinatorial Continuous Information theory Probability theory and statistics Inherent uncertainty (e.g., noise; prediction confidence)... B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 9 / 48
Outline 1 Introduction 2 Methods Machine learning Bayesian networks Directional statistics 3 Applications Introduction to neuroscience Neuron classification Morphological simulation Soma and spines DRCMST Other applications 4 Future work B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 10 / 48
Underpinning: conditional independence Many random variables: intractable distributions 20 binary variables mean 2 20 1 parameters in the joint distribution Fortunately, some variables are sometimes independent of others E.g., if I know that it is very warm, then knowing that it is summer might not make it more likely that many people will be on the beach Factor a joint distribution into smaller local ones P(X 1, X 2, X 3..., X n ) = P(X 1 )P(X 2 X 1 )P(X 3,..., X n X 1, X 2 ) B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 11 / 48
Representation Directed acyclic graph Nodes = variables Arcs encode conditional independencies A local distribution for each parents values combination P(x) = n i=1 P(x i pa(x i )) Can greatly reduce number of parameters http://commons.wikimedia.org/wiki/file:simplebayesnet.svg B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 12 / 48
Inference B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 13 / 48
Inference B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 14 / 48
Some research topics Learning from data NP-hard in the general case Conditional-independence tests Structure scoring (optimization) Inference NP-complete in the general case Exact Approximate Classifiers Specialized learning algorithms Non-standard local probability distributions Hybrid networks Mixtures of polynomials Directional variables B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 15 / 48
Outline 1 Introduction 2 Methods Machine learning Bayesian networks Directional statistics 3 Applications Introduction to neuroscience Neuron classification Morphological simulation Soma and spines DRCMST Other applications 4 Future work B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 16 / 48
Directional statistics Deal with directions, axes, rotations Cannot be studied as regular real-valued variables. Periodicity Real world data: Wind, animal behaviour, neuroscience,... B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 17 / 48
Representation and methods Different ways to represent directional data Directional probability distributions B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 18 / 48
Research topics in CIG Bayesian networks Different local distributions Multi-dimensional classifiers Learning classifiers Big Data... Heuristic optimization Multi-objetive Estimation of distribution algorithms (probabilistic evolutionary) Applications Neuroscience Scientometrics Bioinformatics B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 19 / 48
Outline 1 Introduction 2 Methods Machine learning Bayesian networks Directional statistics 3 Applications Introduction to neuroscience Neuron classification Morphological simulation Soma and spines DRCMST Other applications 4 Future work B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 20 / 48
Projects and collaborations Projects Collaborations Companies B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 21 / 48
Outline 1 Introduction 2 Methods Machine learning Bayesian networks Directional statistics 3 Applications Introduction to neuroscience Neuron classification Morphological simulation Soma and spines DRCMST Other applications 4 Future work B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 22 / 48
The brain Scientific study of the nervous system. Molecular and cellular neuroscience We do not study the brain at macro level (yet)...... but on a micro scale: Neurons 100 billion neurons in the brain 180.000 kilometers of wiring (myelinated white fibers) B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 23 / 48
Neurons Three main parts: Soma, dendrites and axon Neurons communicate with each other using electro-chemical signals Significant differences between neurons B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 24 / 48
Outline 1 Introduction 2 Methods Machine learning Bayesian networks Directional statistics 3 Applications Introduction to neuroscience Neuron classification Morphological simulation Soma and spines DRCMST Other applications 4 Future work B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 25 / 48
Gardener Classification There is an accepted catalogue of neuron types and names But lack of a consistent terminology Every neuroanatomist has is own classification scheme B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 26 / 48
Towards a consensus in naming Learning from the experts Gather data from 42 experts Learn a model (Bayesian network) for each expert B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 27 / 48
Towards a consensus in naming Differences among experts Six clusters of experts (Bayesian network clustering) B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 28 / 48
Outline 1 Introduction 2 Methods Machine learning Bayesian networks Directional statistics 3 Applications Introduction to neuroscience Neuron classification Morphological simulation Soma and spines DRCMST Other applications 4 Future work B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 29 / 48
Morphological simulation Denditric trees Why so different denditric tree shapes? Determine interconnectivity and functional roles B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 30 / 48
Morphological simulation Variables More than 40 variables Evidence and construction variables B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 31 / 48
Morphological simulation B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 32 / 48
Outline 1 Introduction 2 Methods Machine learning Bayesian networks Directional statistics 3 Applications Introduction to neuroscience Neuron classification Morphological simulation Soma and spines DRCMST Other applications 4 Future work B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 33 / 48
Soma spatial characterization Descriptors based on the level curves of a level set function Hybrid Gaussian and angular Bayesian network B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 34 / 48
Spines Related with brain functions like learning and memory 3D active contours to repair fragmented spines Hybrid spatial DBN B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 35 / 48
Outline 1 Introduction 2 Methods Machine learning Bayesian networks Directional statistics 3 Applications Introduction to neuroscience Neuron classification Morphological simulation Soma and spines DRCMST Other applications 4 Future work B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 36 / 48
Main idea Degree-constrained minimum spanning tree Degree constraints Restrict the role of the nodes in the tree to root, intermediate or leaf node Novel permutation-based representation to encode forests of DRCMST B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 37 / 48
Example 20 points where we are interested in building a forest of three trees B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 38 / 48
Application to Neuroscience Applied to optimal neuronal wiring B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 39 / 48
Outline 1 Introduction 2 Methods Machine learning Bayesian networks Directional statistics 3 Applications Introduction to neuroscience Neuron classification Morphological simulation Soma and spines DRCMST Other applications 4 Future work B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 40 / 48
Medical applications Medical decision support systems: Neonatal jaundice treatment B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 41 / 48
Other applications DNA microarray analysis Immunology Alzheimer Parkinson B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 42 / 48
Outline 1 Introduction 2 Methods Machine learning Bayesian networks Directional statistics 3 Applications Introduction to neuroscience Neuron classification Morphological simulation Soma and spines DRCMST Other applications 4 Future work B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 43 / 48
Integration B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 44 / 48
BN & Big data B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 45 / 48
Big data in neuroscience Functional Magnetic Resonance Imaging (fmri) Single Photon Emission Computed Tomography (SPECT) B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 46 / 48
Contact us! Summer School 2015 Computational Intelligence Group bmihaljevic@fi.upm.es luis.rodriguezl@alumnos.upm.es http://cig.fi.upm.es B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 47 / 48
Machine learning in neuroscience Bojan Mihaljevic, Luis Rodriguez-Lujan Computational Intelligence Group School of Computer Science, Technical University of Madrid 2015 IEEE Iberian Student Branch Congress April 24 th, Madrid B. Mihaljevic, L. Rodriguez-Lujan (UPM) Apr, 2015 48 / 48